Tuning the structure and parameters of a neural network by using hybrid Taguchi-genetic algorithm

Area of Science:

• Computational Intelligence
• Machine Learning
• Artificial Neural Networks

Background:

• Tuning feedforward neural network structure and parameters is complex due to numerous parameters and local optima.
• Traditional genetic algorithms (TGA) offer global exploration but can be enhanced for efficiency.
• The Taguchi method provides systematic reasoning for optimizing offspring selection.

Purpose of the Study:

• To introduce and evaluate a hybrid Taguchi-genetic algorithm (HTGA) for optimizing feedforward neural network architecture and weights.
• To demonstrate the HTGA's effectiveness in improving convergence speed and robustness compared to existing methods.

Main Methods:

• A hybrid Taguchi-genetic algorithm (HTGA) is developed by integrating the Taguchi method into the crossover and mutation stages of a traditional genetic algorithm (TGA).
• The Taguchi method's systematic reasoning is used to select superior genes during crossover operations.
• The HTGA is tested on global numerical optimization problems and applied to sunspot number forecasting, associative memory tuning, and the XOR problem.

Main Results:

• The HTGA approach demonstrated superior performance and faster convergence on benchmark optimization problems.
• Application to neural network tuning resulted in partially connected networks, reducing implementation costs.
• The HTGA achieved better results than existing methods for the tested neural network tuning tasks.

Conclusions:

• The hybrid Taguchi-genetic algorithm (HTGA) is a statistically sound and robust method for tuning feedforward neural networks.
• HTGA effectively optimizes both network structure and parameters, leading to more efficient and cost-effective neural network implementations.
• The proposed approach offers significant advantages over traditional genetic algorithms and other existing methods in complex optimization scenarios.